IN DEVELOPMENT · GATED PREVIEW

PRISM

The sanitised LLM gateway.

PRISM sits between your agents and every language model — cloud or local. Every prompt is scanned for PII, developer secrets, and prompt injection before it leaves your perimeter, while responses are logged with live spend tracking.

Private Cluster Development ReleaseActive Evaluation

Automated NER redaction models and JIT proxy endpoints are undergoing private enterprise security audits. Gated access is provisioned on request.

Request Gated Access

Dual-Layer Redaction↓Prompt Injection Defense↓lvls Key Leasing↓Budget Enforcement↓

PRISM · LLM GATEWAY

ACTIVE

Live RedactionDual-layer: NER + Pattern

John Smith→[NAME]

0 items redacted this session0 blocked

OWASP LLM Top 10 Defense

LLM01Blocked

Prompt Injection

LLM02Sanitized

Insecure Output

LLM06Redacted

Info Disclosure

LLM08Scoped

Excessive Agency

Protection Layers

PII Redaction

Active

Prompt Injection

47 blocked

Budget Enforcement

Per-call key acquisition from lvlsNo keys stored in PRISM

PII & Secret Redaction

Dual-Layer Redaction Pipeline

Raw agent prompts never reach public cloud APIs unmodified. PRISM strips confidential customer data and developer credentials before requests cross the cluster perimeter, replacing them with deterministic pseudonyms.

Inbound Agent Request (Raw)UNSAFE

“Process refund for user John Doe (SSN: 482-19-4821, Card: 4111-2948-2941-8492). Connect to internal database via postgres://admin:secret99@192.0.2.100:5432”

Sanitised Outbound Payload (Sent to Model)SANITISED

“Process refund for user <TOKEN_USER_01> (SSN: <TOKEN_SSN_01>, Card: <TOKEN_CC_01>). Connect to internal database via <TOKEN_DB_URI_01>”

Neural NERLayer 1: Protegrity Neural NER

Customer & Personal Identifiers (PII)

Social Security Numbers, Credit Cards, Medical Record Numbers, Full Names, Home Addresses

Pattern RegexLayer 2: Deterministic Pattern Engine

Developer Secrets & Credentials

AWS Access Keys, GitHub PATs, JWT Tokens, SSH Keys, Database Passwords

Perimeter ShieldLayer 2: Infrastructure Redactor

Internal Network Infrastructure

Internal IP addresses (10.x, 192.168.x), Cluster Node Hostnames, VPC Subnets

IP GuardLayer 1 + 2: Semantic Scanner

Proprietary Trade Secrets

Proprietary Algorithm Weights, Customer Tenant IDs, Internal Financial Projections

Adversarial Defense

Prompt Injection & Jailbreak Defense

Untrusted web scrapes and external payloads often contain adversarial prompts designed to hijack autonomous agents. PRISM scans prompts prior to model dispatch.

Instruction Overrides

Flags phrases such as “Ignore previous instructions”, “System override”, and persona switching attacks before token ingestion.

Encoded Payloads

Decodes and evaluates Base64, hex, and unicode obfuscated payloads concealed inside scraped web pages or PDF documents.

Zero Token Wastage

Flagged malicious requests are blocked immediately at the gateway layer (HTTP 403), consuming zero third-party model tokens.

Ephemeral Secret Delivery

Zero Stored Provider Keys

PRISM stores zero OpenAI, Anthropic, or Gemini API keys in its environment or on disk. When an LLM request arrives, PRISM requests a Just-In-Time lease from lvls. The key is delivered via ML-KEM-768 post-quantum transport, used in RAM for the single request, and purged immediately.

Per-Call LeaseKeys exist in memory only for the millisecond duration of the HTTP call.

Zero Disk PersistenceIf the PRISM container is inspected or breached, no provider keys exist on disk.

Post-Quantum TransportAll leased keys delivered via NIST FIPS 203 ML-KEM-768 encryption.

Financial Guardrails

Real-Time Budget & Spend Enforcement

Preventing runaway autonomous loops. PRISM enforces hard monetary spend ceilings per task, per agent, and per day. Requests are blocked before overspend can occur.

Live USD Metering

Calculates input, output, and cached prompt token costs dynamically per provider rate card.

80% Soft Alert

Broadcasts high-priority webhook alert to 4Context and Slack before thresholds are breached.

100% Hard Stop

Subsequent model requests blocked automatically until manual operator budget expansion.

Supported Models & Local FallbackModel Agnostic Routing

AnthropicClaude 3.5 Sonnet · Claude 3 Opus

Cloud·FastGATED ACCESS

OpenAIGPT-4o · GPT-4o-mini · o1

Cloud·FastGATED ACCESS

Google DeepMindGemini 1.5 Pro · Gemini 1.5 Flash

Cloud·FastGATED ACCESS

Self-Hosted ClusterLlama 3 70B · Mistral Large · DeepSeek

On-Premise·Ultra-Low LatencyBENCHMARKING

Explore the full platform

Five integrated components. One unified system.

All components Request Gated Access